ay level
method
defined
best set
atrix by
1. This
ween 2
le (d)
uted by
that can
rmation
., 1996,
is et al.,
ariation
, which
ractice,
s shape
aps the
ons Ri,
yramid,
Stephane Paquis
CBA
Figure 4: Example of simplified road texture images obtained for each surface category (white pixels = Objects, black
pixels = Background and grey pixels = Transitions).
The technique used for the selection of the threshold value t consists of evaluating the distribution of the cooccurrence
matrix coefficients through a function, called measure. Principle lies in considering that Foreground/Background interac-
tion is important when intensities of pairs of pixels are very different. Such information is located in block B3 merged in
matrix. We apply 3 measures (Chanda and Majumder, 1988) on such a block to extract 3 possible threshold values. By
taking into account min and max values, a new matrix mapping is obtained as shown in figure 3.(b). We can create a new
image representation for level /, by assigning a label to each matrix regions.
This procedure is applied along our pyramidal structure and a new one is created, whose levels are quantified on 3 values
corresponding to Object, Background and Transition. Simplified texture version is computed by using a top/down process,
whose construction rule is :
1. If (Father = Object)
e If (Son = Object) then Son = Object
e If (Son = Background) then Son = Transition
e If (Son = Transition) then Son = Object
n3
If (Father 2 Background or Transition) then Nothing
Segmentation results are given in figure 4. A structural analysis of these segmented images is presented in the following
section.
5 EXPERIMENTALS RESULTS
5.1 Procedure of classification
Procedure of classification is divided into 2 steps. Preclassification step consists in distinguishing SD/CBA class from
PA/UTBA one. If we consider road texture from an aggregate arrangement point of view, we can note that only coarse
aggregates appear on PA/UTBA surfaces, besides, background is visible partially. SD/CBA surfaces are composed by
mixture of coarse and fine aggregates and background region is approximatively connex.
Distinction between these 2 categories is performed by analysing 2 components texture images extracted from the sim-
plified texture. Objects number, N 0, is valued on high binary image obtained by affecting to each Transition pixels, the
Objects label. Background number, Np, is extracted from low binary image obtained by assigning to each Transition
pixels the Background label. Separation is done by comparing No /Ng fraction to 1 :
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 689